Deep learning approaches to predicting Autism Spectrum Disorder diagnosis from video data

Abstract

Autism Spectrum Disorder (ASD), a neurodevelopmental condition, affects approximately 1 in 100 individuals worldwide. Confirming a clinical diagnosis of ASD relies predominantly on interviews and questionnaires, yet these approaches have inherent limitations. Presently, the rapid development of Artificial Intelligence (AI) technologies across various domains, including medical research, has spurred considerable interest among researchers exploring AI applications in ASD studies. A noteworthy approach is the utilization of video-based ASD diagnosis with AI, offering advantages in terms of accessibility and information volume compared to other data modalities, such as facial and brain images. In this study, we conducted a search for video and AI-based ASD diagnosis studies published between 2018 and 2024, identifying nine pertinent papers. Our analysis and discussion of these papers, segregated by input features, 1. Facial features, 2. Pose and gait features and 3. Multimodal features. These input features resulted in a promising ASD prediction accuracy on the test data range of 79.7-96.39%. However, we also highlighted certain issues and areas for improvement like out-of-cohort validation, sample size, the black box problem with AI, low specificity, and the establishment of robust and easy video-capturing protocol. These insights contribute valuable information for future clinical applications in the domain of ASD diagnosis.